Rethinking Training Objective For Self-Supervised Monocular Depth Estimation: Semantic Cues To Rescue